Beyond Stateless AI-Models: Exploring Vector Databases for Stateful Knowledge Representation

In 1690, the English philosopher John Locke developed and popularized an ancient old concept called tabula rasa or blank slate. It refers to the philosophical concept that suggests individuals are born with an essentially empty mind. According to this concept, people acquire all their knowledge and traits through experiences and interactions with the world around them. In other words, the tabula rasa mind is stateless; it needs to interact with the world using basic cognitive structures in space and time.

Most machine learning models we use today for knowledge representation are stateless too. Although they are not blank, they interacted with the world during training by ingesting text, audio, depth, heath (maps), IMU, and other datatypes to create a stateless snapshot representation of the world, expressed in space, using vector embeddings. In harmony with the stateless models, vector databases are the shepherds of the time element. Using the models to generate vector embeddings to represent new knowledge, and injecting them in the prompt, we can give generative models access to unique, personal, or proprietary data giving state to the new wave of AI-native applications.

In this talk, we will look at the next evolution in the world of open-source vector databases, where we tighten the relationship of knowledge management in space (the models) and time (the database) by looking at concepts such as Generative Search, Generative Feedback Loops, Near Real-Time Vector Reindexing, and more.